Atria Dawn is an official release, and the full weights are still coming soon

The Atria Dawn site says the model officially launches, “from preview to official release.” That sentence is in the site bundle whose path includes atria-dawn-20261016. The page prints no clock time. The GitHub repository was created on 12 September 2026, and the default branch was pushed on 16 October 2026 at 03:09 UTC. The README and the Chinese README call it an agent model from the Shanghai Artificial Intelligence Laboratory, built on a 744-billion-parameter mixture-of-experts GLM-5.2 base. The 744B figure is the lab’s.

The lab’s about page describes 上海人工智能实验室 as a new type of research institution in China’s AI field, with an address in Xuhui, Shanghai. The project site’s contact line is a Fudan University email.

Pandaily, 17 October 2026 at 07:20 UTC, reports a full release about a month after a mid-September preview. The preview repo internlm/Atria-Dawn-Preview was created on Hugging Face on 11 September 2026. Pandaily’s 512K figure is also on the lab’s API docs. The download table still says 256K.

Weights

The README table labels Atria Dawn and Atria Dawn FP8 as the full release, both at 256K context, both “Coming soon” on Hugging Face and ModelScope. The preview and preview FP8 rows are also 256K, and those are the rows with links. The preview repo is public, ungated, and tagged license:mit and safetensors. A matching ModelScope preview is linked from the table. A Hub search under internlm on 17 October returned the preview repos and no full Atria-Dawn repo. A direct Hub read of internlm/Atria-Dawn returned no model card.

The GitHub tree is the READMEs, a LICENSE, images, and atria-dawn-preview.pdf. No weight files. The LICENSE is MIT, copyright 2026 atria-asi. The README says the code and the model weights are under that MIT license. The full-release weights are still coming soon, so this is not an open-weights release of the full model. The API name in the README is Atria-Dawn.

256K on GitHub, 512K on the API docs

Atria’s API docs list Atria-Dawn at a 512K context window. The page says 512K covers input and generated text in one request, and that the output cap is an integer from 1 to 65,536. A Codex example there sets context_window to 512000 and says not to exceed the model’s real ability. Input modalities in that example are text only.

The same README that prints 256K tells Kimi Code users to set max_context_size to 256000 and calls that “the real limit,” with max output 65536, because it says Atria rejects a larger max_tokens. The project site bundle has the launch line and the benchmark notes. It does not contain the string 512K. The README and the API docs disagree. This account does not pick a winner.

The README says the model accepts text only. The API docs agree. A Codex image attach is described as failing because the model is not multimodal.

Scores on the project table

The README says its table matches the website. The scores are the lab’s, and they were not rerun here. Comparison columns are named DeepSeek V4 Pro 0813, KIMI K3, Qwen 3.8 Max, GLM 5.3, GPT 5.6 sol, and Claude Opus 5. Full release against preview:

Benchmark	Atria Dawn	Preview
JobBench	57.5	50.3
WideSearch	84.6	81.9
PaperBench	89.6	87.0
AutomationBench	54.7	53.8
SWE-bench Pro	61.7	59.6
BrowseComp	91.3	92.5

BrowseComp is a row where the preview is higher. The JobBench, WideSearch, and PaperBench gaps versus preview are 7.2, 2.7, and 2.6 points on the lab’s numbers. On JobBench the same table prints GLM 5.3 at 58.2 and Claude Opus 5 at 68.0, both above 57.5.

Footnotes name the harness. JobBench uses OpenCode on 65 tasks, with Grok-4.3 as judge. PaperBench uses Claude Code for non-GPT models and o3-mini as judge. SWE-bench Pro uses OpenHands v1.23.0. The BrowseComp note describes an in-house ReAct harness with a 256K window. That 256K is the harness setting, a third figure beside the README row and the API docs.

Calling it

International access is the API console and the docs. The China console in the README is discovery.intern-ai.org.cn. The docs show Chat Completions, Messages, and Responses, all with model id Atria-Dawn. The docs page opened here printed no price.

Local serving is listed as SGLang v0.5.13.post1 or newer and vLLM v0.23.0 or newer. The links are a GLM-5.2 SGLang cookbook, which redirected to docs.sglang.io, and a GLM-5.2 vLLM recipe. The README also has a Codex block for https://api.atria-asi.ai/v1 with wire_api set to responses. A “How to Run Locally” section is still a comment placeholder. Until the full weights leave “Coming soon,” the hosted API is the release a caller can use.